I am using keras to train a 1D CNN with time-series as input data to perform a binary classification. The model part of the code is the following: modelo = Sequential() modelo.add(Conv1D(filters=32, kernel_size=5, activation='relu', input_shape=(max_size, 1))) modelo.add(BatchNormalization()) modelo.add(Conv1D(filters=32, kernel_size=5, activation='relu')) modelo.add(BatchNormalization()) modelo.add(Dropout(0.3)) modelo.add(Flatten()) modelo.add(Dense(64, activation='relu')) modelo.add(Dropout(0.3)) modelo.add(Dense(1, activation='sigmoid')) modelo.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) early_stopping = callbacks.EarlyStopping(monitor='loss', patience=1, restore_best_weights=True) modelo.fit(X, etiquetas, epochs=100, batch_size=5, callbacks=[early_stopping]) What I don't understand is that, when the fit function is executed, 9 epochs are computed, all of them with an accuracy higher than 80%. Then, due to early_stopping, training stops at the end of epoch 9. The resulting accuracy is only 54%. Why is there a so high discrepancy between the training accuracies during the epochs and the accuracy printed at the end? I would expect to get a final accuracy around the average of the training epochs accuracy. At least that is the case when I use an LSTM or an MLP. Fortunately, I get very good results with the LSTM and MLP. This means that I don't need the CNN to perform good. However, I would like to know why the CNN does not perform so good or if there is something wrong. Continue reading...
